DeepL taps Nvidia B200 chips to translate the web in 18 days

How long would it take to translate the entire internet? For DeepL, the answer used to be more than half a year. Now, it takes just about two weeks.

The German translation startup announced on Wednesday that it is deploying Nvidia's latest hardware systems, a move that slashes its time to translate the entire web from 194 days down to just 18 days.

Packing 36 Blackwell chips into a rack

DeepL, valued at $2 billion, has long held its ground against Google Translate by building dedicated AI models focused solely on language.

The massive speed boost comes courtesy of Nvidia's DGX SuperPOD system. DeepL is running server racks packed with 36 B200 Grace Blackwell Superchips apiece, tapping into the chipmaker's latest flagship silicon to both train and run giant language architectures.

Stefan Mesken, chief scientist at DeepL, told CNBC that the primary goal is handing vastly more compute to researchers so they can build more advanced models.

The extra horsepower is already feeding into user-facing products. Mesken pointed to Clarify, an interactive feature DeepL launched this year that asks clarifying questions when a word or phrase depends heavily on missing context. That kind of back-and-forth reasoning simply was not technically feasible until this latest generation of hardware arrived, he noted.

Why it matters

For everyday users and global teams, translation tools are finally crossing the bridge from rigid pattern matching to context-aware reasoning. Interactive features like Clarify usually fall apart in real-world use because latency and compute costs make real-time clarification too expensive. As hardware throughput jumps by an order of magnitude, nuanced and domain-specific translations will feel instantaneous.

For the broader AI ecosystem, this marks a clear shift in where cutting-edge chips are landing. High-end hardware like the B200 is no longer just funneling into hyperscalers like Microsoft and Amazon for raw frontier model training. Specialized application builders are now writing big checks for top-tier silicon to protect their product moats.

We think DeepL's focus is sharp: rather than wasting capital chasing general intelligence, it is pouring extreme compute into a single, high-frequency workflow. The lingering challenge, however, will be keeping that edge alive as frontier LLMs become cheaper and inherently better at cross-lingual nuance over time.